Relevant Experience & Education Highlights

Brings 8.7 years of hands-on expertise in computer vision, multimodal learning, video understanding, and production model deployment, strongly aligning with the demands of designing and deploying advanced CV models at Metropolis.

1. Design, develop, and deploy advanced computer vision models for real-world applications, including object detection, tracking, OCR, image search, and scene understanding

Tackled core computer vision challenges in production environments, delivering measurable improvements in accuracy and efficiency.

• Fine-tuned 7B parameter vision-language models on proprietary video datasets at a major streaming service division of a large technology company, boosting precision by 70% for movie scene content descriptor prediction.
• Developed object detection models using YOLO algorithm with PyTorch at a startup, enabling accurate identification of stationery items.
• Built deep learning models with TensorFlow and Keras for artist classification and style prediction from image patches at an AI company serving the art market.
• Achieved 7% improvement over state-of-the-art on internal long-form video datasets and 2% on ActivityNet through scalable models handling up to 3-minute videos at a major technology company.

2. Explore and integrate multi-modal approaches, leveraging visual, textual, and other data modalities for robust solutions

Integrated vision, language, and multimodal data to solve complex real-world problems like content moderation and product matching.

• Productionized multi-level profane content detection combining LLMs and model ensembles for textual and visual modalities at a major streaming service, reducing operator review time by 75% and achieving 90% accuracy across multiple territories.
• Conducted representation learning for joint embedding of text sequences, images, and videos in story-illustration tasks at a research lab within a top-tier university.
• Performed unsupervised product normalization using image and text-based models, increasing coverage by 4% across large catalogs at an e-commerce product platform.

3. Build and optimize deep learning models, ensuring high accuracy, performance, and scalability for deployment in production environments

Optimized large-scale models for domain-specific tasks, enabling production deployment with significant performance gains.

• Instruction-fine-tuned vision-language models for maturity rating and content descriptor prediction on Prime Video datasets at a major technology company.
• Implemented weakly supervised instance segmentation for video understanding, with work published at AMLC 2024.
• Led full SDLC for deep learning projects, including model training and deployment, while training teams in recommendation systems at a startup.

Requirements & Candidate Alignment